计算机科学与探索 ›› 2020, Vol. 14 ›› Issue (10): 1744-1753.DOI: 10.3778/j.issn.1673-9418.1911024

• 图形图像 • 上一篇    下一篇

多尺度非局部注意力网络的小目标检测算法

梁延禹,李金宝   

  1. 1. 黑龙江大学 计算机科学技术学院,哈尔滨 150080
    2. 黑龙江大学 黑龙江省数据库与并行计算重点实验室,哈尔滨 150080
    3. 黑龙江大学 软件学院,哈尔滨 150080
  • 出版日期:2020-10-01 发布日期:2020-10-12

Small Objects Detection Method Based on Multi-scale Non-local Attention Network

LIANG Yanyu, LI Jinbao   

  1. 1. College of Computer Science and Technology, Heilongjiang University, Harbin 150080, China
    2. Key Laboratory of Database and Parallel Computing of Heilongjiang Province, Heilongjiang University, Harbin 150080, China
    3. Software Technology Institute, Heilongjiang University, Harbin 150080, China
  • Online:2020-10-01 Published:2020-10-12

摘要:

现有的小目标检测方法通常采用多尺度特征图或利用多尺度融合特征进行检测,这些方法主要利用了特征图的空间信息而忽略了通道间的相互依赖关系。提出一种新的小目标检测网络,该网络在浅层利用非局部通道注意力模块整合特征的全局空间信息,进而对通道间的信息进行校准。从空间域及通道域获取特征的远距离依赖关系,增强浅层特征中小目标的上下文语义信息。同时,通过密集连接结构增强深层部分的特征提取能力,获取丰富的目标信息,提高目标检测任务的准确率和实时性。实验结果表明,该算法在PASCAL VOC、MS COCO数据集中得到了较好的检测结果,并且在保证检测速度的前提下,能有效提高小目标的检测准确率。

关键词: 目标检测, 卷积神经网络(CNN), 非局部注意力, 密集连接

Abstract:

Existing small objects detection methods usually detect small objects by multi-scale feature maps or multi-scale fusion features. However, these methods mainly utilize the spatial information of the feature maps but ignore the interdependencies between channels. This paper proposes a novel small objects detection network, the network uses the non-local channel attention module to integrate the global spatial information of the features in the shallow layer, and then calibrates the information between channels. And from the spatial domain and the channel domain, it obtains the long-distance dependence of features and enhances the contextual semantic information of small targets in the shallow features. In addition, the network enhances the feature extraction capability of the deep features through dense connection structure, obtains rich target information, and improves the accuracy and real-time of the object detection task. The experimental results show that the algorithm has good detection results on PASCAL VOC and MS COCO datasets, and it can effectively improve the detection accuracy of small objects under the premise of ensuring the detection speed.

Key words: object detection, convolutional neural network (CNN), non-local attention, dense connection